LmRaC: a functionally extensible tool for LLM interrogation of user experimental results

Area of Science:

• Artificial Intelligence
• Bioinformatics
• Scientific Research Tools

Background:

• Large Language Models (LLMs) show promise but face challenges with accuracy and source fabrication in scientific research.
• Retrieval-augmented generation (RAG) enhances LLMs by providing access to external data for improved reasoning and source traceability.

Purpose of the Study:

• To introduce LmRaC, an LLM-based tool designed to answer complex scientific questions within the context of user-specific experimental results.
• To enable dynamic knowledge base construction from scientific literature for controlled information retrieval.

Main Methods:

• LmRaC utilizes RAG to build domain-specific knowledge bases (RAGdom) from PubMed sources.
• It answers questions using only the RAG knowledge base with paragraph-level citations, preventing hallucination.
• LmRaC integrates user experimental data (RAGexp), including quantitative results and protocols, to answer specific experimental questions via a REST API (RAGfun).

Main Results:

• LmRaC effectively answers complex scientific questions by grounding responses in user-provided data and curated literature.
• The system virtually eliminates hallucination and fabrication by strictly adhering to sourced information.
• Quantitative experimental data can be queried and analyzed within the LmRaC framework.

Conclusions:

• LmRaC offers a robust solution for leveraging LLMs in scientific research by ensuring accuracy, completeness, and authoritativeness.
• The tool facilitates deeper insights into experimental results by integrating diverse data sources and providing traceable answers.